The Identifiability of Approximate Vocal Tract Impulse Response Magnitudes.

摘要

It is shown that the vocal tract impulse response magnitude should be less variable for a given speaker than other acoustic measures of his speech. Cepstrum analysis is used to deconvolve the vocal tract impulse response and glottal pressure wave of each of 1850 speech segments taken from running English speech. Linear correlation coefficients derived from pairs of impulse responses are shown to differ, depending upon whether the two impulse responses were taken from the same speaker's utterances, from speakers of the same sex and/or vocal history, or from altogether different speakers. A feasible method of speaker identification which can in principle operate automatically is developed from this approach and tested. (Author)